Abstract: In order to improve the prediction precision of calibration model,the near infrared spectroscopy(NIR) calibration model for the starch in tobacco was studied.Spectral variables were chosen by means of full spectrum(FS) and variance spectrum(VS),spectral wavelength by means of genetic algorithm(GA),via partial least squares,the calibration models,FS+PLS,VS+PLS and GA+PLS,were established,and the starch contents in 100 flue-cured tobacco samples were predicted.The results showed that:1) R_c² and root mean squares error of cross validation(RMSECV) of the three models,FS+PLS(1 557 variables),VS+PLS(781 variables) and GA+PLS(72 variables) were 0.976 4 and 0.433,0.987 1 and 0.332,0.988 5 and 0.314,respectively.2) Comparing with FS+PLS and VS+PLS,the number of variables in GA+PLS was 4.62% and 9.22% of that in FS and VS,its main factor number reduced from 15 to 12,R_c² increased from 0.976 4 to 0.988 5,and RMSECV decreased from 0.433 to 0.314.3) The predicted results of starch contents in 100 tobacco samples indicated that,R_p² and RMSEP of FS+PLS,VS+PLS and GA+PLS models were 0.965 2 and 0.780,0.984 3 and 0.501,0.985 3 and 0.496,respectively.The paired T test between the prediction values and the test values was performed,the Sig.value,T value and average relative error(%) were 0.271,1.107 and 17.48% for FS+PLS,0.973,0.034 and 13.13% for VS+PLS,0.722,0.357 and 13.12% for GA+PLS,respectively.There were no significant differences between the results predicted by the three methods and the corresponding test values.The RSD values of FS+PLS,VS+PLS and GA+PLS were 10.34%,6.98% and 4.76%,respectively.The prediction precision of GA+PLS model was better than that of FS+PLS and VS+PLS models.This method provides a reference for improving the precision of prediction model for a complex chemical system.